CN110597628A

CN110597628A - Model distribution method and device, computer readable medium and electronic equipment

Info

Publication number: CN110597628A
Application number: CN201910812937.1A
Authority: CN
Inventors: 王星雅
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2019-08-29
Filing date: 2019-08-29
Publication date: 2019-12-20
Anticipated expiration: 2039-08-29
Also published as: CN110597628B

Abstract

The embodiment of the application provides a distribution method and device of an image analysis model, a computer readable medium and electronic equipment. The method is executed by an analysis node in a block chain, and comprises the following steps: acquiring image sample data provided by at least one data node in a block chain network; training a first image analysis model on an analysis node by using image sample data; and synchronizing the parameters of the trained second image analysis model to at least one node outside the analysis nodes in the blockchain network, so that after the parameters of the second image analysis model are configured to a deployed third image analysis model by each node, image analysis is performed by using the third image analysis model with configured parameters, wherein the model architecture of the third image analysis model is the same as that of the second image analysis model. According to the technical scheme, more image data can be acquired, the training effect of the image analysis model is improved, the synchronization of model parameters can be realized, and the updating efficiency of the image analysis model is improved.

Description

Model distribution method and device, computer readable medium and electronic equipment

Technical Field

The present application relates to the field of machine learning technologies, and in particular, to a method and an apparatus for distributing an image analysis model, a computer-readable medium, and an electronic device.

Background

With the development of artificial intelligence, various machine learning models such as a deep neural network and the like are rapidly iterated in the field. In terms of computer vision, more image data is often needed to train a better image analysis model, so how to acquire more image data capable of being used for training the image analysis model has become a common challenge in the industry; in addition, because the training of the image analysis model consumes a large amount of computing power, and the running of the image analysis model uses relatively little computing power, the device used for training the image analysis model and the device using the image analysis model are often different, which results in that the image analysis model on the application device may lag behind the newly trained model and cannot meet the latest application requirement.

Disclosure of Invention

Embodiments of the present application provide a method and an apparatus for distributing an image analysis model, a computer-readable medium, and an electronic device, so as to improve the number of acquired image data used for training the image analysis model at least to a certain extent, and improve the distribution efficiency of the image analysis model.

Other features and advantages of the present application will be apparent from the following detailed description, or may be learned by practice of the application.

According to an aspect of the embodiments of the present application, there is provided a method for distributing an image analysis model, the method being performed by an analysis node in a blockchain network, the method including: acquiring image sample data provided by at least one data node in a block chain network, wherein the block chain network comprises a plurality of nodes; training a first image analysis model on the analysis node by using the image sample data; and synchronizing the parameters of the trained second image analysis model to at least one node except the analysis nodes in the block chain network, so that after the parameters of the second image analysis model are configured to a third image analysis model deployed on each node, each node performs image analysis by using the third image analysis model with configured parameters, wherein the model architecture of the third image analysis model is the same as that of the second image analysis model.

In some embodiments of the present application, based on the foregoing solution, the synchronizing the trained parameters of the second image analysis model to at least one node other than the analysis node in the blockchain network includes:

determining the number of image sample data provided by each data node;

determining a synchronization sequence to each data node in the block chain network according to the number;

and synchronizing the parameters of the trained second image analysis model to the data nodes in the block chain network according to the synchronization sequence.

In some embodiments of the present application, based on the foregoing solution, the determining, according to the number, a synchronization order to the data nodes in the blockchain network includes:

and sequencing the data nodes from large to small according to the number, wherein the sequencing is used as a synchronous sequence of the data nodes in the block chain network.

determining the number of image sample data provided by each data node;

determining the data nodes with the number larger than a first preset number threshold value as target data nodes;

and synchronizing the parameters of the trained second image analysis model to each target data node in the block chain network.

In some embodiments of the present application, based on the foregoing solution, after obtaining image sample data provided by at least one data node in the blockchain network, the method further includes:

and if the number of the obtained image sample data which is not used for training is less than a second preset number threshold, re-obtaining the image sample data provided by at least one data node in the block chain network until the number of the obtained image sample data which is not used for training reaches the second preset number threshold.

According to an aspect of the embodiments of the present application, there is provided an apparatus for distributing an image analysis model, the apparatus being an analysis node in a blockchain network, the apparatus including: the system comprises a data acquisition module, a data acquisition module and a data processing module, wherein the data acquisition module is configured to acquire image sample data provided by at least one data node in a block chain network, and the block chain network comprises a plurality of nodes; a model training module configured to train a first image analysis model on the analysis node using the image sample data; and the synchronization module is configured to synchronize the trained parameters of the second image analysis model to at least one node except the analysis nodes in the blockchain network, so that after the parameters of the second image analysis model are configured to a third image analysis model deployed on each node, each node performs image analysis by using the third image analysis model with configured parameters, wherein the model architecture of the third image analysis model is the same as that of the second image analysis model.

According to an aspect of embodiments of the present application, there is provided a computer-readable medium on which a computer program is stored, which, when executed by a processor, implements a distribution method of image analysis models as described in the above embodiments.

According to an aspect of an embodiment of the present application, there is provided an electronic device including: one or more processors; a storage device for storing one or more programs which, when executed by the one or more processors, cause the one or more processors to implement a method of distributing an image analysis model as described in the above embodiments.

In the technical solutions provided in some embodiments of the present application, on one hand, data for training an image analysis model is obtained by using a block chain network, and as the data source is wider, more image data can be obtained, so that the training effect of the image analysis model can be improved; on the other hand, the parameters of the image analysis model obtained through training are synchronized to other nodes in the block chain network, so that the newly trained image analysis model can be rapidly applied to other nodes, and the updating efficiency of the image analysis model is improved.

It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the application.

Drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and together with the description, serve to explain the principles of the application. It is obvious that the drawings in the following description are only some embodiments of the application, and that for a person skilled in the art, other drawings can be derived from them without inventive effort. In the drawings:

FIG. 1 shows a network architecture diagram of a distribution method application of an image analysis model according to one embodiment of the present application;

FIG. 2A shows a schematic interface diagram for monitoring a source of image sample data for training an image analysis model when the distribution method of the image analysis model is applied according to an embodiment of the present application;

FIG. 2B illustrates a schematic interface diagram of the interface displayed to summarize the source of image sample data when the summarization button of FIG. 2A is triggered according to one embodiment of the present application;

fig. 3 is a schematic diagram of a network architecture in which a method for distributing an image analysis model according to an embodiment of the present application is applied in the medical imaging field;

FIG. 4 shows a flow diagram of a method of distribution of an image analysis model according to an embodiment of the present application;

FIG. 5 shows a detailed flowchart of step S440 of FIG. 4 according to an embodiment of the present application;

FIG. 6 shows a detailed flowchart of step S420 in FIG. 4 according to an embodiment of the present application;

FIG. 7 shows a detailed flowchart of steps following step S420 and step 440 of FIG. 6 according to an embodiment of the present application;

FIG. 8 shows a network architecture diagram of a distribution method application of an image analysis model according to another embodiment of the present application;

FIG. 9 is a schematic diagram illustrating a structure of an abnormal medical image data detection model based on a generative confrontation network according to an embodiment of the present application;

FIG. 10 shows a flow diagram of a training process for an abnormal medical image data detection model according to an embodiment of the present application;

FIG. 11 shows a block diagram of a distribution apparatus of image analysis models according to an embodiment of the present application;

FIG. 12 illustrates a schematic structural diagram of a computer system suitable for use in implementing the electronic device of an embodiment of the present application.

Detailed Description

Example embodiments will now be described more fully with reference to the accompanying drawings. Example embodiments may, however, be embodied in many different forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the concept of example embodiments to those skilled in the art.

Furthermore, the described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to give a thorough understanding of embodiments of the application. One skilled in the relevant art will recognize, however, that the subject matter of the present application can be practiced without one or more of the specific details, or with other methods, components, devices, steps, and so forth. In other instances, well-known methods, devices, implementations, or operations have not been shown or described in detail to avoid obscuring aspects of the application.

The block diagrams shown in the figures are functional entities only and do not necessarily correspond to physically separate entities. I.e. these functional entities may be implemented in the form of software, or in one or more hardware modules or integrated circuits, or in different networks and/or processor means and/or microcontroller means.

The flow charts shown in the drawings are merely illustrative and do not necessarily include all of the contents and operations/steps, nor do they necessarily have to be performed in the order described. For example, some operations/steps may be decomposed, and some operations/steps may be combined or partially combined, so that the actual execution sequence may be changed according to the actual situation.

The application firstly provides a distribution method of an image analysis model. The image analysis model is a model for obtaining a corresponding analysis or processing result according to an input of an image, and can be established based on model architectures in various computer vision and image processing fields, for example, models of a convolutional neural network model, a deep learning model, a reinforcement learning model and the like, and the analysis or processing result of the image analysis model can be various types of results selected according to needs. Distribution of image analysis models refers to the process of synchronizing image analysis models from one device to one or more other terminals or devices.

Fig. 1 shows a network architecture diagram of a distribution method application of an image analysis model according to an embodiment of the present application.

As shown in fig. 1, the network architecture is a blockchain network, and the blockchain network includes a plurality of nodes, i.e., an analysis node 110 and a data node 120 and 150, respectively. The nodes are abstracted devices or device clusters, each node may adopt various devices having functions of storing data, processing data, and communicating with external devices, such as a server, a desktop computer, a server cluster, and the like, and the same terminal device may be used between the analysis nodes and the data nodes and between the data nodes, or different terminal devices may be used between the analysis nodes and the data nodes. In this embodiment, the analysis node 110 is an execution terminal of a distribution method of image analysis models, the analysis node 110 and each data node are both deployed with image analysis models, and model architectures of the image analysis models deployed on the nodes are the same, but parameters of the image analysis models deployed on the analysis node and each data node may be different, and generally, the parameters of the image analysis model on the analysis node are superior to or equal to the parameters of the image analysis model deployed on each data node, that is, the performance of the image analysis model on the analysis node is superior to or equal to the image analysis model deployed on each data node; at least one data node in each data node provides image sample data for training of an image analysis model on the analysis node, the image sample data can be transmitted to the data node from other nodes or submitted to the data node by a user of the data node, the analysis node can synchronize parameters of the image analysis model to one or more data nodes after training the image analysis model to a certain degree by using the obtained image sample data, so that the data nodes directly use the synchronized parameters to run the image analysis model which is locally deployed on each data node, timely update of the image analysis model on each data node receiving the parameters is realized, and the data nodes can rapidly apply the updated image analysis model to meet the latest application requirements.

Fig. 2A illustrates an interface diagram used for monitoring a source of image sample data used for training an image analysis model when the distribution method of the image analysis model is applied according to an embodiment of the present application. The interface may be a Web page such as HTML5, a client interface, an App (Application program) interface, an applet interface, or the like, data displayed on the interface may be stored by the analysis node, the interface may be displayed on the analysis node of the blockchain network, or the analysis node may transmit the data to other nodes according to requests of other nodes in the blockchain network, so that the data is displayed on other nodes in the blockchain network in the form of an interface shown in fig. 2A. Referring to fig. 2A, the interface shows statistics of the number of image sample data provided by three data nodes to the analysis node for training the image analysis model in 1 month 1 day, 1 month 2 day, and 1 month 3 day, respectively, where the number of image sample data provided by a first data node in 1 month 2 day is 480, the number of image sample data provided in 1 month 3 day is 950, the number of image sample data provided by a second data node in 1 month 1 day is 800, the number of image sample data provided in 1 month 3 day is 750, and for a third data node, the number of image sample data provided in 1 month 1 day is 380, the number of image sample data provided in 1 month 2 day is 700, and the number of image sample data provided in 1 month 3 day is 500. The content displayed by the interface can be clear that the quantity of the image sample data provided by each data node to the analysis node is not necessarily fixed and can change with time, the quantity of the image sample data provided by each data node on the same day is different, and each data node does not provide the image sample data to the analysis node every day, for example, the first data node does not provide the image sample data to the analysis node on 1 month and 2 days, so that the source data node of the image sample data acquired by the analysis node every day or each time can be unfixed.

Through the interface shown in fig. 2A, the user can intuitively know the quantity of the image sample data provided by each data node to the analysis node every day.

FIG. 2B illustrates a schematic interface diagram of the interface displayed to summarize the sources of image sample data when the summarization button of FIG. 2A is triggered according to one embodiment of the present application. The "summarization" button is shown in fig. 2A, and the button of the interface shown in fig. 2A may be triggered in various ways, such as by clicking a mouse, and when the summarization button is triggered, the data in the interface shown in fig. 2A is summarized, and the interface shown in fig. 2B is displayed. The interface shown in fig. 2B shows statistics of the number of image sample data provided by three data nodes to the analysis node for training the image analysis model in 1 st day 1 to 1 st day 3, and for the first data node, since the number of image sample data provided by the first data node in 1 st day 2 and 1 st day 3 is 480 and 950 in fig. 2A, respectively, the interface shown in fig. 2B shows that the number of image sample data provided by the first data node in 1 st day 1 to 1 st day 3 is 1430(═ 480+ 950); for the second data node, since the numbers of image sample data provided by the second data node on 1 month, 1 day, and 1 month, 3 days are 800 and 750, respectively, in fig. 2A, the interface shown in fig. 2B displays that the number of image sample data provided by the second data node on 1 month, 1 day, to 1 month, 3 days is 1550(═ 800+ 750); for the third data node, since the numbers of image sample data provided by the third data node on days 1, 2 and 3 are 380, 700 and 500 respectively in fig. 2A, the interface shown in fig. 2B displays that the number of image sample data provided by the third data node on days 1, 1 and 3 is 1580 (380 +700+500), so that it can be calculated that the total number of image sample data for training the image analysis model obtained by the analysis node on days 1, 1 and 3 is 4560 (1430+1550+ 1580).

Through the interface shown in fig. 2B, the user can macroscopically and integrally know the quantity and scale of the image sample data provided by each data node to the analysis node within a time period, so that a guarantee is provided for the user to evaluate the training progress of the image analysis model on the analysis node.

Fig. 3 is a schematic diagram of a network architecture in which a method for distributing an image analysis model according to an embodiment of the present application is applied in the medical imaging field.

The medical image is generally an image that can be used in the medical field to determine the health and development status of a human, and may be, for example, a bone image, a chest X-ray image, or the like. Medical images can provide important information for judging the health and development conditions of human beings, for example, chest X-ray images can be used for judging whether a person has pneumonia.

Referring to fig. 3, the network architecture includes an analysis node 320, a first medical institution node 330, a second medical institution node 340, a first storage node 350, and a second storage node 360, which are connected by a blockchain 310, and each node can send data to other nodes or receive data from other nodes through the blockchain 310. The storage node may provide image sample data to the analysis node 320 in various cooperative ways with the medical facility node. For example, in one aspect, the first medical institution node 330 may send image sample data to be provided to the analysis node 320 to the second medical institution node 340, the second medical institution node 340 verifies the correctness of the image sample data, sends the verified image sample data to the second storage node 360, and sends the obtained image sample data to the analysis node 320 by the second storage node 360; on the other hand, the second medical institution node 340 transmits image sample data to be provided to the analysis node 320 to the first medical institution node 330, the first medical institution node 330 verifies the correctness of the image sample data, the image sample data passing the verification is transmitted to the first storage node 350, and the first storage node 350 transmits the obtained image sample data to the analysis node 320. Therefore, each image sample data obtained by the analysis node 320 can be verified by two medical institution nodes, so that the training effect of the image analysis model on the analysis node can be improved, and the performance of the trained image analysis model is further improved.

FIG. 4 shows a flow diagram of a method of distribution of an image analysis model according to an embodiment of the present application. The distribution method of the image analysis model provided by the embodiment is executed by an analysis node in a blockchain network. The analysis node in the blockchain network may specifically be various types of terminal devices, such as a server, a desktop computer, and the like. Referring to fig. 4, the method for distributing the image analysis model at least includes step S410, step S420 and step S440, and the following is described in detail:

in step S410, image sample data provided by at least one data node in the blockchain network is acquired.

The blockchain network includes a plurality of nodes. I.e. the blockchain network comprises at least an analysis node and one data node.

The data node may be any device with data storage and data transmission capabilities. For example, the data nodes may be database servers, and the types of devices used by the data nodes may be the same or different, and the data nodes may be the same or different as devices actually used by the analysis node.

The blockchain network is a novel application mode of computer technologies such as distributed data storage, point-to-point transmission, a consensus mechanism and an encryption algorithm, and can be any one of a public chain, a private chain and a alliance chain.

The image sample data is data which can be used for training of an image analysis model, and may include image data and corresponding labels, and the types of the image sample data may be various, and may be selected according to application requirements. For example, in the field of garbage classification, image sample data may include photos of garbage and garbage types corresponding to the photos; for another example, in the field of intelligent medical treatment, the image sample data may include an image of a lesion and a lesion recognition result corresponding to the image.

In an embodiment of the present application, the acquiring image sample data provided by at least one data node in a blockchain network includes:

and acquiring image sample data provided by at least one data node in the block chain network every a preset time period.

The method has the advantages that the image sample data are obtained regularly for training the image analysis model, and the image sample data are not obtained immediately every time the data node just obtains the image sample data, so that the resource consumption caused by frequently obtaining the data for many times is reduced, and the data obtaining cost is reduced to a certain extent.

acquiring image sample data sent by at least one data node in a block chain network.

The method and the device have the advantages that once the data node obtains the image sample data, the obtained image sample data can be immediately sent to the analysis node, so that the analysis node can obtain the latest image sample data in time, and the updating speed of the image analysis model is improved.

In an embodiment of the present application, the acquiring image sample data sent by at least one data node in a blockchain network includes:

authenticating at least one data node in the blockchain network through a communication connection established with the at least one data node in the blockchain network;

and acquiring image sample data sent by the data node passing the authentication.

The authentication may be performed in various ways, such as authentication using a digital certificate encrypted based on a public-private key, or authentication using a specific security rule or mechanism.

The method and the device have the advantages that the legality of the acquired image sample data is improved and the safety of the block chain network is ensured by only acquiring the image sample data sent by the data node passing the authentication.

and pulling the image sample data to at least one data node in the block chain network.

The advantage of this embodiment is, through drawing image sample data to data node voluntarily according to the needs of analysis node, can just get data voluntarily when needing data for data turnover efficiency on the analysis node promotes greatly, thereby has reduced the data storage expense on the analysis node to a certain extent.

In an embodiment of the present application, the pulling image sample data to at least one data node in a blockchain network includes:

acquiring the storage space occupancy rate and the memory utilization rate of the analysis node;

and under the condition that the storage space occupancy rate and the memory utilization rate of the analysis node meet preset conditions, pulling image sample data to at least one data node in the block chain network.

The method has the advantages that the data node is removed to pull the image sample data only when the storage space occupancy rate and the memory utilization rate of the analysis node meet the preset conditions, so that the action of removing the data node to pull the image sample data also depends on the storage space occupancy rate and the memory utilization rate of the analysis node, and the analysis node can be ensured to stably run when the data node is removed from the analysis node to pull the data.

In an embodiment of the present application, the predetermined condition is that the storage space occupancy of the analysis node is less than a predetermined storage space occupancy threshold and the memory usage rate is less than a predetermined memory usage rate threshold.

The method has the advantages that the data node is only required to pull the image sample data when the storage space occupancy rate and the memory utilization rate of the analysis node are both sufficiently small, and the operation reliability and stability of the analysis node are guaranteed.

In an embodiment of the application, the predetermined condition is that a weighted sum of the storage space occupancy rate and the memory usage rate of the analysis node is determined to be smaller than a predetermined weighted sum threshold according to a preset weight.

The manner of obtaining the storage space occupancy rate and the memory usage rate of the analysis node may be various, for example, the storage space occupancy rate and the memory usage rate are obtained by crawling the analysis node through a preset script.

and pulling image sample data from the data node corresponding to the time node in the data node and time node corresponding relation table when the current time is equal to one time node in the data node and time node corresponding relation table according to a preset data node and time node corresponding relation table.

The advantage of this embodiment is that, by setting a time node for each data node individually, and then pulling data to the corresponding data node at each time node, image sample data does not need to be pulled to a plurality of data nodes dispersedly when image sample data is acquired, and the efficiency of pulling image sample data is improved to a certain extent.

In an embodiment of the present application, after obtaining image sample data provided by at least one data node in the blockchain network, the method further includes:

The method has the advantages that the acquisition of the image sample data in the current round is stopped only when the number of the acquired image sample data which are not used for training reaches a certain number, and then the acquired image sample data can be used for training the image analysis model, so that the centralized training of the image analysis model is realized, and the utilization rate of resources used by the model training is improved.

In step S420, a first image analysis model on the analysis node is trained using the image sample data.

As described above, the image analysis model may be various, and may be, for example, a convolutional neural network model, a deep learning model, a reinforcement learning model, or a combination of one or more models.

The image analysis model can output a model of a corresponding judgment or processing result in accordance with input of image data.

The first image analysis model may be a model applied to various fields, and the image sample data for training the first image analysis model is image sample data of a corresponding field. For example, in the bone age recognition field, the first image analysis model may be a bone age recognition model, and the image sample data for training the first image analysis model may be hand bone image data. For another example, in the field of garbage classification, the first image analysis model may be a garbage classification model, and the image sample data for training the first image analysis model may be garbage image data.

The image sample data may include an image and a corresponding tag, where the tag corresponds to a result that the first image analysis model should output, so that supervised learning of the first image analysis model may be implemented, for example, in the field of garbage classification, the tag corresponding to the garbage image is a category to which the garbage image belongs, and the categories may be kitchen garbage, harmful garbage, recyclable materials, and the like; in the field of bone age identification, the label corresponding to the hand bone image is the bone age of the hand bone image.

The training of the first image analysis model is a process of determining parameters of the first image analysis model, and the first image analysis model may be trained by various methods, for example, the first image analysis model may be generally trained by a SGD (Stochastic gradient descent) algorithm.

In order to solve the problems that when an image analysis model is trained in the field of medical imaging, abnormal images are difficult to acquire and when abnormal image data is detected only by using rules, the coverage range is limited due to the fact that all the rules cannot be exhausted, the embodiment of the application further provides the following solution.

In an embodiment of the present application, the image sample data includes medical image data and a corresponding tag, the tag indicates that the corresponding medical image data is abnormal or normal, fig. 6 shows a specific flowchart of step S420 in fig. 4 according to an embodiment of the present application, referring to fig. 6, step S420 may include the following steps:

in step S420', a first image analysis model on the analysis node is trained using normal medical image data in the image sample data, where the first image analysis model is an abnormal medical image data detection model based on a generative countermeasure network, and the normal medical image data is medical image data whose label indicates that the corresponding medical image data is normal.

The medical image may be various forms of images in the medical field, including but not limited to image data of the human body such as bones, fundus, lungs, brain, etc.

The normal medical image data, i.e., the corresponding label, shows that the person who generated the medical image data does not have a disease that can be characterized by the medical image data, and correspondingly, the abnormal medical image data, i.e., the corresponding label, shows that the person who generated the medical image data does not have a disease that can be characterized by the medical image data. For example, if the medical image data is brain image data, the normal medical image data may be brain image data determined not to have a brain tumor, and the abnormal medical image data may be brain image data determined to have a brain tumor.

Generative Adaptive Networks (GAN) is a deep learning model that passes through at least two modules in a framework: the mutual game learning of the Generative Model (Generative Model) and the Discriminative Model (Discriminative Model) yields a reasonably good output.

Therefore, the abnormal medical image data detection model based on the generative countermeasure network is used for detecting whether the medical image data shows that the corresponding individual has a certain disease or shows a certain symptom, i.e., detecting whether the medical image data is abnormal. It should be noted that, in the present embodiment, only the normal medical image data is used when the abnormal medical image data detection model based on the generative countermeasure network is trained.

In an embodiment of the present application, based on the embodiment shown in fig. 6, the blockchain network further includes an expert verification node, where the expert verification node is configured to verify whether a tag corresponding to the medical image data is accurate, and after acquiring image sample data provided by at least one data node in the blockchain network, the method further includes:

randomly selecting a preset number of groups of medical image data and corresponding labels from the medical image data and the corresponding labels provided by each data node, and sending the medical image data and the corresponding labels to the expert verification node so as to obtain a verification result of the expert verification node for the corresponding labels of the medical image data;

determining the label estimation accuracy of each data node according to the verification result;

acquiring medical image data provided by data nodes with label estimation accuracy rate larger than a preset label estimation accuracy rate threshold value as target medical image data;

training a first image analysis model on the analysis node using the image sample data, comprising:

and training a first image analysis model on the analysis node by using the target medical image data and the corresponding label.

Compared with a common data node, the expert verification node is more authoritative in determining the legality of the label corresponding to the medical image data, and can provide more accurate labels for the medical image data. For example, the label of the medical image data provided by the data node may be marked by a general physician in a county level (district level) or a city level hospital, and the expert verification node may be provided in a provincial level hospital and be responsible for verification by the expert physician in the corresponding field. The verification result may be various information that can be used to determine the tag estimation accuracy, for example, the verification result may be a determination result of whether the tag corresponding to each group of medical image data is correct or not, or may be a tag labeled by the expert verification node for each group of medical image data, and therefore, the manner of determining the tag estimation accuracy of each data node according to the verification result may be various, for example, if the verification result is a determination result of whether the tag corresponding to each group of medical image data is correct or not, the tag estimation accuracy of one data node may be equal to a ratio of a correct group number to a predetermined number of groups of medical image data corresponding to the data node and corresponding tags, and a determination result of the corresponding tag is a ratio of the correct group number to the predetermined number.

In this embodiment, for a data node, not all medical image data and corresponding tags provided by the data node are sent to the expert verification node, but only a predetermined number of groups of medical image data and corresponding tags selected from the data node are sent to the expert verification node for verification. Therefore, for a data node, the group number of the medical image data and the corresponding labels selected and sent to the expert verification node for verification is not enough to accurately indicate the label accuracy of the data node, but at the same time, since the medical image data and the corresponding labels sent to the expert verification node are randomly selected, the accuracy determined according to the randomly selected data can be considered to be approximately equal to the label accuracy, and therefore, the accuracy determined according to the randomly selected data is referred to as the label estimation accuracy, which means the estimation value of the label accuracy.

In this embodiment, on one hand, the medical image data provided by which data nodes and the corresponding labels can be used for training the first image analysis model is determined according to the label estimation accuracy, so that the data quality for training the first image analysis model is ensured, and the training effect of the model is improved; on the other hand, the label estimation accuracy is determined by only using a part of data provided by each data node, so that the workload of the expert verification node is reduced, and the time for training the first image analysis model is also improved.

The embodiment shown in fig. 6 has the advantages that, because the number of abnormal medical image data is usually small, that is, the number of individuals suffering from diseases or presenting certain symptoms is small, the requirement for training the abnormal medical image data detection model cannot be met only by the abnormal medical image data, while the number of normal medical image data is usually much larger than that of the abnormal medical image data, in this embodiment, the abnormal medical image data detection model is detected by the normal medical image data, so that the condition that the number of the abnormal medical image data cannot meet the requirement of the training model is avoided; in addition, when abnormal medical image data is detected, a method generally adopted can be a method of judging by using a rule set by an expert, but all rules capable of detecting the abnormal medical image data cannot be exhausted at present, and the coverage range of the detection is limited, so that the problem that the coverage of the current rules is limited can be solved.

In an embodiment of the present application, based on the embodiment shown in fig. 6, the blockchain network includes a plurality of sub-networks, where each sub-network includes at least one data node and at least one medical institution node, and the image sample data provided by each data node is synchronized by the medical institution node to the data node in the sub-network to which the medical institution node belongs after the medical institution node obtains the image sample data.

The method has the advantages that the image sample data of each medical institution node in the same sub-network is uniformly managed by the data nodes under the sub-network, so that the separation of data acquisition and data storage is realized, each data node only provides the image sample data in the sub-network to which the data node belongs, the storage load of the data node is reduced, and the acquisition efficiency of the data node is improved.

In an embodiment of the present application, based on the foregoing embodiment, the synchronizing the trained parameters of the second image analysis model to at least one node other than the analysis node in the blockchain network includes:

and synchronizing the parameters of the trained second image analysis model to each medical institution node in the block chain network.

In the embodiment, the parameters of the trained second image analysis model are synchronized to the medical institution nodes requiring the parameters in the block chain network, so that the medical institution nodes can update the model in time, and the timeliness of model update is ensured for the nodes requiring the application of the abnormal medical image data detection model.

Fig. 8 shows a network architecture diagram of a distribution method application of an image analysis model according to another embodiment of the present application. As shown in fig. 8, the network architecture includes an analysis node 810, a first sub-network 820, a second sub-network 830, a third sub-network 840 and a fourth sub-network 850, wherein each sub-network includes a data node and at least one medical institution node, and bidirectional data transmission can be performed between the data node in one sub-network and the data nodes in the other sub-networks and between the data node in the sub-network and the analysis node. In fig. 8, the circle included in each sub-network and connected to the data node represents a medical institution node in the sub-network, for example, the medical institution node 821 is one of the medical institution nodes belonging to the first sub-network 820, and it can be seen that each sub-network includes at least one medical institution node, but the number of medical institution nodes included in each sub-network is not fixed, for example, two medical institution nodes are included in the third sub-network 840 and the fourth sub-network 850, but three medical institution nodes are included in the first sub-network 820 and the second sub-network 830, respectively.

When the distribution method of the image analysis model is applied to the network architecture shown in fig. 8, a specific process may be such that: each medical institution node in each sub-network generates medical image data by using medical equipment and prints a corresponding label, then each medical structure node sends the medical image data and the corresponding label to the data node in the sub-network to which each medical structure node belongs, each data node stores the medical image data and the corresponding label and forwards the medical image data and the corresponding label to an analysis node, and the analysis node divides the medical image data into normal medical image data and abnormal medical image data according to the label corresponding to each medical image data; then, the analysis node trains a first image analysis model on the analysis node by using the normal medical image data, and tests a second image analysis model which is trained by using the abnormal medical image data; finally, if the second image analysis model passes the test, the analysis node synchronizes the parameters of the second image analysis model to at least one data node, the data node obtaining the parameters sends the parameters to each medical institution node in the sub-network to which the data node belongs, and each medical institution node can configure the obtained parameters to a third image analysis model already deployed on each medical institution node, so that each medical institution node can perform image analysis by using the third image analysis model configured with the parameters, that is, perform the detection of abnormal medical image data.

In an embodiment of the present application, based on the embodiment shown in fig. 6, the abnormal medical image data detection model based on the generative countermeasure network includes: an original encoder, a decoder, a re-encoder and a classifier, wherein:

the original encoder is used for extracting a first implicit vector of the original medical image data after receiving the input original medical image data;

the decoder is used for reconstructing medical image construction data corresponding to the original medical image data according to the first implicit vector output by the original encoder;

the secondary encoder is used for extracting a second implicit vector of the medical image construction data after the medical image construction data output by the decoder is obtained so as to train and test the abnormal medical image data detection model based on the generative countermeasure network;

the classifier is used for judging whether the original medical image data and the medical image construction data output by the decoder are abnormal or not.

The model structure of the general generative countermeasure network (GAN) includes only a Generator (Generator) and a Discriminator (Discriminator), and the abnormal medical image data detection model based on the generative countermeasure network used in the present embodiment includes four parts, namely, an Original Encoder (Original Encoder), a Decoder (Decoder), a secondary Encoder (second Encoder) and a classifier (Discriminator), and the model structure is designed to not only train the abnormal medical image data detection model by using the normal medical image data, but also realize the high-precision detection of the abnormal medical image data.

In an embodiment of the present application, the structure of the re-encoder is consistent with the structure of the original encoder, the original encoder includes a plurality of convolution layers, a batch normalization layer connected to at least one convolution layer, and a linear rectification activation function layer with leakage connected to at least one batch normalization layer, the decoder includes a plurality of inverse convolution layers, a linear rectification activation function layer with leakage connected to at least one inverse convolution layer, a linear rectification activation function layer with leakage connected to at least one linear rectification activation function layer with leakage, and an output activation function layer located after the last inverse convolution layer, and the classifier includes at least one convolution layer and a classification layer.

Since the structure of the re-encoder is consistent with that of the original encoder, the type of the result output by the re-encoder and the original encoder is the same, and corresponding feature representation, namely, implicit vector, is output according to the input of the medical image data. In the present embodiment, the original encoder, the decoder, the re-encoder, and the classifier all include a multi-layer structure, so each component in the abnormal medical image data detection model based on the generative countermeasure network provided in the present embodiment is a layer-by-layer structure.

In this embodiment, a convolution (Conv) layer, i.e. a convolutional neural network layer, may be used to extract a feature representation of the input data; a batch normalization (BatchNorm) layer may be used to keep the inputs to each layer of the neural network the same during training of the model; the leakage-containing Linear rectification activation function (Leaky Rectified Linear Unit, Leaky ReLU) layer is an activation function, is a variant of the Linear rectification activation function (Rectified Linear Unit, ReLU), and can avoid neuron death; a deconvolution (convTranspose) layer is an inverse process of convolution and can amplify input feature data; the output activation function layer may map the input data to a non-linear space, and may be based on various activation functions, such as a Sigmoid function, a Tanh function, and the like; the classification layer may be configured to output a corresponding classification result according to input data, and may be implemented based on a Softmax function, for example.

In an embodiment of the present application, based on the above-described embodiment, a structure or framework of the abnormal medical image data detection model based on the generative countermeasure network may be as shown in fig. 9.

Fig. 9 is a schematic structural diagram of an abnormal medical image data detection model based on a generative countermeasure network according to an embodiment of the present application. Referring to fig. 9, the structure of the abnormal medical image data detection model includes: an encoding module 920, a countermeasure module 930, and a classification module 940, the encoding module 920 may include an original encoder, the countermeasure module 930 may include a decoder and a re-encoder, the classification module 940 may include a classifier, the input raw medical image data 910 is sent to the encoding module 920 and the classification module 940 respectively for performing the training of the abnormal medical image data detection model, after the original medical image data 910 is sent to the encoding module 920, the original encoder in the encoding module 920 extracts the first implicit vector of the original medical image data 910, and after the first implicit vector is sent to the countermeasure module 930, the decoder in the countermeasure module 930 would construct the medical image construction data using the first implicit vector, the re-encoder in the countermeasure module 930 may extract and output the second implicit vector 950 of the medical image construction data output by the decoder; the classifier in the classification module 940 may obtain the raw medical image data 910 and the medical image configuration data correspondingly output by the decoder in the countermeasure module 930 according to the input of the raw medical image data 910, and correspondingly output the detection result 960 of the raw medical image data.

For the embodiment shown in fig. 9, the original encoder, decoder, re-encoder, and classifier may all be in a layer-by-layer stacked configuration. The original encoder may comprise a five-layer structure, each layer structure comprising one convolutional layer, wherein the first layer structure may comprise only one convolutional layer, the second layer structure may comprise one convolutional layer and a leaky linear rectifying activation function layer connected to the convolutional layer, and for the third to five-layer structures, the specific structure contained in each layer structure may be the same, for example, each layer structure may comprise one convolutional layer, a batch normalization layer connected to the convolutional layer, and a leaky linear rectifying activation function layer connected to the batch normalization layer; the decoder may comprise a five-layer structure, each layer structure may comprise one deconvolution layer, wherein the first layer structure may comprise one deconvolution layer and a leaky linear rectification activation function layer connected to the deconvolution layer, and the specific structure contained in each layer structure may be the same for the second to fourth layer structures, for example, each layer structure may comprise one deconvolution layer, a leaky linear rectification activation function layer connected to the deconvolution layer, and a linear rectification activation function layer connected to the leaky linear rectification activation function layer, and the fifth layer structure may comprise one deconvolution layer and an output activation function layer connected to the deconvolution layer; the structure of the re-encoder can be the same as that of the original encoder, and is not described herein again; the classifier may include a six-layer structure, wherein the first to five-layer structure may be the same as the five-layer structure of the original encoder or the re-encoder, and the last layer structure is a classification layer.

In an embodiment of the present application, based on the above embodiment, the training a first image analysis model on the analysis node by using normal medical image data in the image sample data includes:

repeatedly executing a training process on the abnormal medical image data detection model on the analysis node by using the normal medical image data in the image sample data until a preset condition is met, wherein the training process comprises the following steps:

the encoder and decoder are trained by minimizing the loss function:

wherein x is the input original medical image data, and z is G_E(x) A first implicit vector output for the original encoder,constructing data for the medical image output by the decoder,a second implicit vector, | x-G, output for the re-encoder_D(z) | is an index for measuring a difference between original medical image data and the medical image configuration data,alpha and beta are weight coefficients for an index used to measure the difference between the first implicit vector and the second implicit vector, the encoder comprising an original encoder and a re-encoder;

the parameters of the encoder and decoder are fixed and the classifier is trained by minimizing the loss function:

wherein d (x) is a classifier function that is a mapping between an input of given medical image data and a result of an output of a last layer of the classifier preceding a classification layer corresponding to the inputThe relationship between the beams is defined as,the index is used for measuring the difference between the result correspondingly output by the middle layer of the classifier after the original medical image data is input into the classifier and the result correspondingly output by the middle layer of the classifier after the data output by the decoder is input into the classifier;

the parameters of the classifier are fixed and the encoder and decoder are trained in a antagonistic manner using the following loss function:

wherein d (x) is a classifier function that is a mapping between an input of given medical image data and an output of a last layer preceding a classification layer in the classifier corresponding to the input,the index is used for measuring the difference between the result correspondingly output by the middle layer of the classifier after the original medical image data is input into the classifier and the result correspondingly output by the middle layer of the classifier after the data output by the decoder is input into the classifier.

Inclusion of | x-G in the loss function due to L1_D(z)‖ and the difference between the original medical image data and the medical image construction data and the difference between the first implicit vector and the second implicit vector are respectively measured, so that the parameters of an original encoder, a decoder and a re-encoder can be optimized through the training of an L1 loss function, and the generated original medical image data and the medical image construction data and the generated first implicit vector and the generated second implicit vector are similar enough on the whole; l2 loss functionThe method comprises the steps that after original medical image data are input into a classifier, the result correspondingly output by the middle layer of the classifier and the result correspondingly output by the middle layer of the classifier after the data output by the decoder are input into the classifier are measured, wherein the middle layer of the classifier can be one of any two layers of the classifier, for example, the middle layer can be the last layer before the classification layer, so that the L2 loss function is trained by fixing the parameters of the encoder and the decoder, and the classification error of the classifier can be reduced; the L3 loss function is trained by fixing the parameters of the classifier to be unchanged, and this process is to train the encoder and decoder in a game manner, for example, in the embodiment of fig. 9, the parameters of the decoder and re-encoder in the countermeasure module 930 are mainly trained, and by increasing the L3 loss function, the classifier cannot distinguish whether the samples are normal samples, so that the performance of the encoder and decoder can be further improved.

In an embodiment of the present application, based on the above embodiment, synchronizing the trained parameters of the second image analysis model to at least one node in the blockchain network, where the node is located outside the analysis nodes, so that after configuring the parameters of the second image analysis model to a third image analysis model deployed on each node, each node performs image analysis using the third image analysis model configured with the parameters, where the method includes:

synchronizing the parameters of the abnormal medical image data detection model obtained through training to at least one node except the analysis node in the block chain network, so that after the parameters of the second image analysis model are configured to the abnormal medical image data detection model deployed on each node, each node detects the medical image data based on the following formula:

wherein ,is a stand forThe second implicit vector output by the re-encoder,is an index used to measure the difference between the first implicit vector and the second implicit vector.

Because the abnormal medical image data detection model only utilizes normal medical image data for training and modeling, after the abnormal medical image data is input into the abnormal medical image data detection model, the abnormal medical image data detection model cannot accurately generate the implicit vector, so that the difference between the first implicit vector and the second implicit vector is caused, and therefore, after one medical image data is input into the abnormal medical image data detection model, if the medical image data is obtained through calculation, the difference between the first implicit vector and the second implicit vector is causedIf the size is large enough, the medical image data can be determined to be abnormal medical image data.

Next, a model training process when the distribution method of the image analysis model provided by the present application is applied to the medical image field and the image analysis model is an abnormal medical image data detection model will be described with reference to fig. 10. Fig. 10 shows a flowchart of a training process of an abnormal medical image data detection model according to an embodiment of the present application. After the analysis node acquires the medical image data, classifying the medical image data into normal medical image data and abnormal medical image data, and then performing model training by using the normal medical image data to obtain a trained abnormal medical image data detection model; then, the abnormal medical image data is used to verify the performance of the trained abnormal medical image data detection model, and the verification modes can be various, for example, the verification can be performed by judging the detection accuracy of the model, when the model is not verified, the model needs to be retrained until the model is verified, and when the model is verified, the medical image data to be recognized can be detected by using the model which is verified, so that whether the medical image data to be recognized is abnormal can be judged.

Continuing to refer to fig. 6, in step S440, synchronizing the trained parameters of the second image analysis model to at least one node in the blockchain network other than the analysis nodes, so that each node performs image analysis using a third image analysis model with configured parameters after configuring the parameters of the second image analysis model to the third image analysis model deployed on each node.

Wherein the third image analysis model has the same model architecture as the second image analysis model.

The second image analysis model is obtained by training the first image analysis model in step S420, so that the first image analysis model and the second image analysis model are the same image analysis model in different training schedules, the model architectures or structures of the first image analysis model and the second image analysis model are the same, and the first image analysis model and the second image analysis model are different in that the parameters of the two image analysis models are different. The parameters of the image analysis model and the model architecture are the core factors determining the performance of the image analysis model.

Before each node configures the parameters of the second image analysis model to the third image analysis model deployed on each node, although the model architecture of the third image analysis model deployed on each node is the same as that of the second image analysis model, the performance of the third image analysis model after the parameters are initialized or applied only lags behind that of the second image analysis model, and the parameters of the third image analysis model deployed on each node can be updated by synchronizing the parameters of the second image analysis model to each node.

Because the model architecture of the third image analysis model is the same as that of the second image analysis model, after the parameters of the second image analysis model synchronized by the analysis nodes are obtained by each node and configured to the third image analysis model deployed on each node, the third image analysis model used by each node is basically the same as that of the second image analysis model, so that the image analysis model trained on the analysis nodes can be timely synchronized to other nodes needing to apply the image analysis model.

In an embodiment of the application, each node is preset with a parameter passing interface, and each node configures parameters of the second image analysis model to a third image analysis model deployed on each node by calling the parameter passing interface on each node.

In an embodiment of the present application, the synchronizing the trained parameters of the second image analysis model to at least one node other than the analysis node in the blockchain network includes:

and synchronizing the parameters of the trained second image analysis model to the data nodes which provide image sample data for the analysis nodes in the block chain network.

The advantage of this embodiment is that, by synchronizing the parameters of the second image analysis model to the data nodes that provide image sample data, the aggressiveness of these data nodes in providing image sample data can be improved to a certain extent, and further the training efficiency of the image analysis model can be improved.

acquiring image sample data and parameter receiving node identification provided by at least one data node in a block chain network;

the synchronizing the trained parameters of the second image analysis model to at least one node in the blockchain network other than the analysis node includes:

and synchronizing the parameters of the trained second image analysis model to the parameter receiving nodes corresponding to the parameter receiving node identifications in the block chain network according to the parameter receiving node identifications.

The parameter receiving node identifier provided by one data node is an identifier of a node to which the data node indicates that the analysis node needs to synchronize the parameters of the second image analysis model after training to obtain the second image analysis model.

The method has the advantages that the data nodes providing the image sample data can customize the nodes to which the parameters of the second image analysis model need to be synchronized, and the training and distribution of the image analysis model of the block chain network are more coordinated and scientific.

In an embodiment of the present application, a table of correspondence between node identifiers and addresses is preset, and synchronizing, according to each parameter receiving node identifier, a parameter of a second image analysis model obtained by training to a parameter receiving node corresponding to each parameter receiving node identifier in the block chain network includes:

determining addresses corresponding to the parameter receiving node identifications by inquiring a node identification and address corresponding relation table;

and synchronizing the parameters of the trained second image analysis model to the nodes corresponding to the addresses in the block chain network according to the addresses.

determining the number of image sample data provided by each data node;

In an embodiment of the application, the determining, according to the number, a synchronization sequence to the data nodes in the blockchain network includes:

The method has the advantages that the parameters of the synchronized second image analysis model can be preferentially obtained by the data nodes with more image sample data, namely the data nodes making more contribution to the training of the image analysis model, and the fairness of parameter synchronization is improved.

dividing the number of image sample data provided by each data node into a preset number of intervals according to the sequence from large to small, wherein the number of each image sample data only belongs to one interval;

sequencing the data nodes from large to small according to the number;

and randomly sequencing the data nodes with the number belonging to the same interval as the synchronous sequence of each data node in the block chain network.

Because the contribution of each data node to the training image analysis model is not necessarily in a positive correlation with the quantity of the provided image sample data, for example, although some data nodes provide a large quantity of image sample data, most data nodes have poor quality, such as image blurring and poor quality of image sample data, the embodiment has the advantage that the parameter synchronization sequence of the data nodes with the same quantity belonging to the same interval is still dependent on the quantity of the image sample data provided by the data nodes, and the fairness of the parameter synchronization is improved to a certain extent.

determining the number of image sample data provided by each data node;

The method has the advantages that the parameters of the second image analysis model can be obtained only when the number of the provided image sample data reaches a certain number of data nodes, and the enthusiasm of the data nodes for providing the image sample data can be improved to a certain extent, so that the training efficiency of the image analysis model can be improved.

In an embodiment of the present application, step S430 may be further included after step S420 in fig. 6, and step S440 may specifically include step S440 ″.

Fig. 7 shows a detailed flowchart of steps following step S420 and step 440 in fig. 6 according to an embodiment of the present application. Referring to fig. 7, the following steps may be included:

in step S430, the trained abnormal medical image data detection model based on the generative countermeasure network is tested by using the abnormal medical image data in the image sample data.

The abnormal medical image data is medical image data of which the label indicates that the corresponding medical image data is abnormal.

The process of testing the abnormal medical image data detection model based on the generative countermeasure network is a process of evaluating whether the performance of the abnormal medical image data detection model meets the expected requirements.

Since the purpose of the abnormal medical image data detection model based on the generative countermeasure network is to detect or identify the abnormal medical image data, the abnormal medical image data detection model can be tested using the abnormal medical image data.

The process of testing the abnormal medical image data detection model may be such that the abnormal medical image data is input to the abnormal medical image data detection model, the detection result output by the abnormal medical image data detection model can be obtained, the detection result may be abnormal or normal, if the detection result is normal, the abnormal medical image data detection model performs error identification on the abnormal medical image data, a plurality of abnormal medical image data can be input to the abnormal medical image data detection model respectively, the detection result corresponding to each abnormal medical image data is obtained, the ratio of the number of abnormal medical image data whose detection result is abnormal to the number of all abnormal medical image data input to the abnormal medical image data detection model is obtained as the accuracy rate of testing the abnormal medical image data detection model, a determination may then be made as to whether the test performed on the abnormal medical image data detection model passed based on a comparison of the accuracy rate to a predetermined accuracy rate threshold.

For example, the accuracy obtained by testing the abnormal medical image data detection model is 70%, and the predetermined accuracy threshold is 80%, and since 70% is less than 80%, it can be determined that the abnormal medical image data detection model fails the test.

In step S440 ″, parameters of the abnormal medical image data detection model based on the generative countermeasure network that passes the test are synchronized to at least one node other than the analysis node in the blockchain network.

After the abnormal medical image data detection model based on the generative countermeasure network passes the test, the model is considered to be capable of operating well and have good performance, and the detection requirement of general abnormal medical image data can be met, so that the parameters of the model can be synchronized to other nodes.

The embodiment shown in fig. 7 has the advantage that in the case that the abnormal medical image data detection model passes the test, the parameters of the abnormal medical image data detection model are synchronized to other nodes, so that after the other nodes configure the obtained parameters on the model, the deployed model of the nodes has good performance, and the detection effect is ensured to a greater extent.

In an embodiment of the present application, each node in the blockchain network other than the analysis node deploys a third image analysis model, fig. 5 shows a specific flowchart of step 440 in fig. 4 according to an embodiment of the present application, and referring to fig. 5, step S440 may include the following steps:

in step S440', the trained parameters of the second image analysis model are synchronized to all nodes except the analysis node in the blockchain network, so that after the parameters of the second image analysis model are configured to the third image analysis model deployed on each node, each node performs image analysis by using the third image analysis model with configured parameters.

In this embodiment, the trained parameters of the second image analysis model are synchronized to all nodes except the analysis node in the block chain network, so that more nodes in the block chain network can update the image analysis model in time, and the application range of the image analysis model is expanded.

In summary, according to the distribution method of the image analysis model provided in the embodiment of fig. 4, on one hand, the data used for training the image analysis model is obtained by using the blockchain network, and as the data source is wider, more image data can be obtained, so that the training effect of the image analysis model can be improved; on the other hand, by synchronizing the parameters of the trained image analysis model to other nodes in the blockchain network, so that the newly trained image analysis model can be rapidly applied to other nodes, the distribution and update efficiency of the image analysis model is improved, therefore, a high-efficiency iteration mechanism of a whole set of image analysis model comprising the links of data acquisition, model training, model distribution and the like of the image analysis model is established, in addition, an analysis node does not need to store a large amount of data for the training image analysis model independently, after the image sample data is acquired and used for training the image analysis model, the data can be deleted at any time, the method and the device have the advantages that the training effect of the image analysis model is guaranteed, and meanwhile, the number of image sample data which are simultaneously stored on the analysis node and used for training the image analysis model is reduced, so that the storage cost is reduced.

The following describes embodiments of the apparatus of the present application, which may be used to perform the distribution method of the image analysis model in the above-described embodiments of the present application. For details that are not disclosed in the embodiments of the apparatus of the present application, please refer to the embodiments of the distribution method of the image analysis model described above in the present application.

Fig. 11 shows a block diagram of a distribution apparatus of an image analysis model according to an embodiment of the present application.

Referring to fig. 11, an apparatus 1100 for distributing an image analysis model according to an embodiment of the present application, the apparatus 1100 being an analysis node in a blockchain network, the apparatus 1100 comprising: a data acquisition module 1110, a model training module 1120, and a synchronization module 1130.

The data obtaining module 1110 is configured to obtain image sample data provided by at least one data node in a block chain network, where the block chain network includes a plurality of nodes; a model training module 1120, configured to train a first image analysis model on the analysis node using the image sample data; a synchronization module 1130, configured to synchronize a parameter of the trained second image analysis model to at least one node other than the analysis node in the block chain network, so that each node configures the parameter of the second image analysis model to a third image analysis model deployed on each node, and then performs image analysis using the third image analysis model configured with the parameter, where the model architecture of the third image analysis model is the same as that of the second image analysis model.

In some embodiments of the present application, based on the foregoing solution, each node in the blockchain network except the analysis node deploys a third image analysis model, and the synchronization module 1130 is further configured to: and synchronizing the parameters of the trained second image analysis model to all nodes except the analysis node in the block chain network.

In some embodiments of the present application, based on the foregoing scheme, the image sample data includes medical image data and a corresponding label, where the label indicates that the corresponding medical image data is abnormal or normal, and the model training module 1120 is further configured to: and training a first image analysis model on the analysis node by using normal medical image data in the image sample data, wherein the first image analysis model is an abnormal medical image data detection model based on a generative countermeasure network, and the normal medical image data is medical image data of which the label indicates that the corresponding medical image data is normal.

In some embodiments of the present application, based on the foregoing scheme, after training the first image analysis model on the analysis node with the image sample data, the model training module 1120 is further configured to: testing the trained abnormal medical image data detection model based on the generative countermeasure network by using abnormal medical image data in the image sample data, wherein the abnormal medical image data is medical image data of which the label indicates that the corresponding medical image data is abnormal; the synchronization module 1130 is further configured to: synchronizing parameters of the generated countermeasure network-based abnormal medical image data detection model that passes testing to at least one node in the blockchain network other than the analysis node.

In some embodiments of the present application, based on the foregoing solution, the blockchain network includes a plurality of sub-networks, where each sub-network includes at least one data node and at least one medical institution node, and the image sample data provided by each data node is synchronized by a medical institution node to the data node in the sub-network to which the medical institution node belongs after the medical institution node obtains the image sample data.

In some embodiments of the present application, based on the foregoing solution, the abnormal medical image data detection model based on generative confrontation network includes: an original encoder, a decoder, a re-encoder and a classifier, wherein:

In some embodiments of the present application, based on the foregoing solution, the structure of the re-encoder is consistent with the structure of the original encoder, the original encoder includes a plurality of convolution layers, a batch normalization layer connected to at least one convolution layer, and a linear rectification activation function layer with leakage connected to at least one batch normalization layer, the decoder includes a plurality of inverse convolution layers, a linear rectification activation function layer with leakage connected to at least one inverse convolution layer, a linear rectification activation function layer connected to at least one linear rectification activation function layer with leakage, and an output activation function layer located after the last inverse convolution layer, and the classifier includes at least one convolution layer and a classification layer.

In some embodiments of the present application, based on the foregoing scheme, the model training module 1120 is further configured to:

the encoder and decoder are trained by minimizing the loss function:

wherein d (x) is a classifier function that is a mapping between an input of given medical image data and an output of a last layer preceding a classification layer in the classifier corresponding to the input,the index is used for measuring the difference between the result correspondingly output by the middle layer of the classifier after the original medical image data is input into the classifier and the result correspondingly output by the middle layer of the classifier after the data output by the decoder is input into the classifier;

wherein d (x) is a classifier function that is a mapping between an input of given medical image data and an output of a last layer preceding a classification layer in the classifier corresponding to the input,the method is used for measuring the result correspondingly output by the middle layer of the classifier and the data output by the decoder after the original medical image data is input into the classifierThe intermediate layers of the classifier correspond to an indicator of the difference between the output results.

In some embodiments of the present application, based on the foregoing, the synchronization module 1130 is further configured to:

wherein ,a second implicit vector output for the re-encoder,is an index used to measure the difference between the first implicit vector and the second implicit vector.

It should be noted that the computer system 1200 of the electronic device shown in fig. 12 is only an example, and should not bring any limitation to the functions and the scope of use of the embodiments of the present application.

As shown in fig. 12, the computer system 1200 includes a Central Processing Unit (CPU)1201, which can perform various appropriate actions and processes, such as performing the methods described in the above embodiments, according to a program stored in a Read-Only Memory (ROM) 1202 or a program loaded from a storage section 1208 into a Random Access Memory (RAM) 1203. In the RAM 1203, various programs and data necessary for system operation are also stored. The CPU 1201, ROM 1202, and RAM 1203 are connected to each other by a bus 1204. An Input/Output (I/O) interface 1205 is also connected to bus 1204.

The following components are connected to the I/O interface 1205: an input section 1206 including a keyboard, a mouse, and the like; an output section 1207 including a Display device such as a Cathode Ray Tube (CRT), a Liquid Crystal Display (LCD), and a speaker; a storage section 1208 including a hard disk and the like; and a communication section 1209 including a network interface card such as a LAN (Local area network) card, a modem, or the like. The communication section 1209 performs communication processing via a network such as the internet. A driver 1210 is also connected to the I/O interface 1205 as needed. A removable medium 1211, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, or the like, is mounted on the drive 1210 as necessary, so that a computer program read out therefrom is mounted into the storage section 1208 as necessary.

In particular, according to embodiments of the application, the processes described above with reference to the flow diagrams may be implemented as computer software programs. For example, embodiments of the present application include a computer program product comprising a computer program embodied on a computer readable medium, the computer program comprising program code for performing the method illustrated by the flow chart. In such an embodiment, the computer program may be downloaded and installed from a network through the communication section 1209, and/or installed from the removable medium 1211. The computer program executes various functions defined in the system of the present application when executed by a Central Processing Unit (CPU) 1201.

It should be noted that the computer readable medium shown in the embodiments of the present application may be a computer readable signal medium or a computer readable storage medium or any combination of the two. A computer readable storage medium may be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the foregoing. More specific examples of the computer readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer diskette, a hard disk, a Random Access Memory (RAM), a Read-Only Memory (ROM), an Erasable Programmable Read-Only Memory (EPROM), a flash Memory, an optical fiber, a portable Compact Disc Read-Only Memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In the present application, a computer readable storage medium may be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device. In this application, however, a computer readable signal medium may include a propagated data signal with computer readable program code embodied therein, for example, in baseband or as part of a carrier wave. Such a propagated data signal may take many forms, including, but not limited to, electro-magnetic, optical, or any suitable combination thereof. A computer readable signal medium may also be any computer readable medium that is not a computer readable storage medium and that can communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device. Program code embodied on a computer readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wired, etc., or any suitable combination of the foregoing.

The flowchart and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present application. Each block in the flowchart or block diagrams may represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams or flowchart illustration, and combinations of blocks in the block diagrams or flowchart illustration, can be implemented by special purpose hardware-based systems which perform the specified functions or acts, or combinations of special purpose hardware and computer instructions.

The units described in the embodiments of the present application may be implemented by software, or may be implemented by hardware, and the described units may also be disposed in a processor. Wherein the names of the elements do not in some way constitute a limitation on the elements themselves.

As an aspect, the present application also provides a computer-readable medium, which may be contained in the electronic device described in the above embodiments; or may exist separately without being assembled into the electronic device. The computer readable medium carries one or more programs which, when executed by an electronic device, cause the electronic device to implement the method described in the above embodiments.

It should be noted that although in the above detailed description several modules or units of the device for action execution are mentioned, such a division is not mandatory. Indeed, the features and functionality of two or more modules or units described above may be embodied in one module or unit, according to embodiments of the application. Conversely, the features and functions of one module or unit described above may be further divided into embodiments by a plurality of modules or units.

Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein may be implemented by software, or by software in combination with necessary hardware. Therefore, the technical solution according to the embodiments of the present application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a usb disk, a removable hard disk, etc.) or on a network, and includes several instructions to enable a computing device (which can be a personal computer, a server, a touch terminal, or a network device, etc.) to execute the method according to the embodiments of the present application.

Other embodiments of the present application will be apparent to those skilled in the art from consideration of the specification and practice of the embodiments disclosed herein. This application is intended to cover any variations, uses, or adaptations of the invention following, in general, the principles of the application and including such departures from the present disclosure as come within known or customary practice within the art to which the invention pertains.

It will be understood that the present application is not limited to the precise arrangements described above and shown in the drawings and that various modifications and changes may be made without departing from the scope thereof. The scope of the application is limited only by the appended claims.

Claims

1. A method of distributing an image analysis model, the method being performed by an analysis node in a blockchain network, the method comprising:

acquiring image sample data provided by at least one data node in a block chain network, wherein the block chain network comprises a plurality of nodes;

training a first image analysis model on the analysis node by using the image sample data;

and synchronizing the parameters of the trained second image analysis model to at least one node except the analysis nodes in the block chain network, so that after the parameters of the second image analysis model are configured to a third image analysis model deployed on each node, each node performs image analysis by using the third image analysis model with configured parameters, wherein the model architecture of the third image analysis model is the same as that of the second image analysis model.

2. The method of claim 1, wherein the image sample data includes medical image data and a corresponding label indicating that the corresponding medical image data is abnormal or normal,

and training a first image analysis model on the analysis node by using normal medical image data in the image sample data, wherein the first image analysis model is an abnormal medical image data detection model based on a generative countermeasure network, and the normal medical image data is medical image data of which the label indicates that the corresponding medical image data is normal.

3. The method of claim 2, wherein after training a first image analysis model on the analysis node with the image sample data, the method further comprises:

testing the trained abnormal medical image data detection model based on the generative countermeasure network by using abnormal medical image data in the image sample data, wherein the abnormal medical image data is medical image data of which the label indicates that the corresponding medical image data is abnormal;

synchronizing parameters of the generated countermeasure network-based abnormal medical image data detection model that passes testing to at least one node in the blockchain network other than the analysis node.

4. The method of claim 2, wherein the blockchain network comprises a plurality of sub-networks, each sub-network comprising at least one data node and at least one medical facility node, wherein the image sample data provided by each data node is synchronized by a medical facility node to the data nodes in the sub-network to which the medical facility node belongs after the medical facility node obtains the image sample data.

5. The method of claim 2, wherein the generative confrontation network-based abnormal medical image data detection model comprises: an original encoder, a decoder, a re-encoder and a classifier, wherein:

6. The method of claim 5, wherein the re-encoder has a structure identical to that of the original encoder, wherein the original encoder comprises a plurality of convolutional layers, a batch normalization layer connected to at least one convolutional layer, and a linear rectifying activation function layer with leakage connected to at least one batch normalization layer, wherein the decoder comprises a plurality of inverse convolutional layers, a linear rectifying activation function layer with leakage connected to at least one inverse convolutional layer, a linear rectifying activation function layer with leakage connected to at least one linear rectifying activation function layer with leakage, and an output activation function layer located after the last inverse convolutional layer, and wherein the classifier comprises at least one convolutional layer and a classification layer.

7. The method of claim 6, wherein training a first image analysis model on the analysis node using normal medical imagery data in the image sample data comprises:

the encoder and decoder are trained by minimizing the loss function:

wherein d (x) is a classifier function that is a mapping between an input of given medical image data and an output of a last layer preceding a classification layer in the classifier corresponding to the input,the method is used for measuring the result correspondingly output by the middle layer of the classifier after the original medical image data is input into the classifier and the resultAfter the data output by the decoder is input into the classifier, the intermediate layer of the classifier correspondingly outputs an index of the difference between the results;

8. The method according to claim 7, wherein the synchronizing the trained parameters of the second image analysis model to at least one node in the blockchain network other than the analysis nodes so that each node performs image analysis using a third image analysis model with configured parameters after configuring the parameters of the second image analysis model to the third image analysis model deployed on each node comprises:

9. The method of claim 1, wherein each node in the blockchain network other than the analysis node deploys a third image analysis model, and the synchronizing the trained parameters of the second image analysis model to at least one node in the blockchain network other than the analysis node comprises:

and synchronizing the parameters of the trained second image analysis model to all nodes except the analysis node in the block chain network.

10. An apparatus for distributing an image analysis model, wherein the apparatus is an analysis node in a blockchain network, and the apparatus comprises:

the system comprises a data acquisition module, a data acquisition module and a data processing module, wherein the data acquisition module is configured to acquire image sample data provided by at least one data node in a block chain network, and the block chain network comprises a plurality of nodes;

a model training module configured to train a first image analysis model on the analysis node using the image sample data;

and the synchronization module is configured to synchronize the trained parameters of the second image analysis model to at least one node except the analysis nodes in the blockchain network, so that after the parameters of the second image analysis model are configured to a third image analysis model deployed on each node, each node performs image analysis by using the third image analysis model with configured parameters, wherein the model architecture of the third image analysis model is the same as that of the second image analysis model.